Elamur Research⌕
☰
home / papers / 1996 / A Corpus Study of Negative Imperatives in Natural Language Instructions

A Corpus Study of Negative Imperatives in Natural Language Instructions

Paper · 1996 · Keith Vander Linden and Barbara Di Eugenio

Summary

The paper looks at how instructional texts warn readers away from wrong actions. The authors call these "preventative expressions" and ask when writers choose a plain negative imperative ("don't X") over a cautionary form such as "be careful not to X". From a roughly 4 MB corpus of written English instructions (recipes, do-it-yourself manuals, game and software help, forms), they pulled candidates with lexical probes, sampled them, and filtered out false hits. This left 239 examples.

Both authors coded each example for one form feature (DONT vs neg-TC) and two functional features:

  • intentionality: does the reader consciously intend the action?
  • awareness: does the reader know the consequences are bad?

They checked reliability with the Kappa statistic and tested form–function correlations with χ². Both features turned out to co-vary with form:

  • DONT goes with actions done deliberately and with readers unaware of the danger.
  • neg-TC goes with accidental actions and with readers who already know the danger.

These correlations were turned into a C4.5 decision tree that selects warning forms in the DRAFTER instruction-generation system. The study shows how pragmatic assumptions about the reader drive surface choice in generation.

Key points

  • The corpus is about 4 MB of written English instructions, including 1.7 MB of recipes and 1.2 MB from two DIY manuals. The probed forms occurred 1,175 times, which is 2.5% of all expressions.
  • About 100 examples were randomly sampled per form. After removing non-imperatives and non-negatives, 239 remained: don't 78, do not 89, take care 17, make sure 3, be careful 46, be sure 6.
  • Raw agreement between coders was 76.1% for intentionality and 92.5% for awareness. The corresponding Kappa values were only 0.51 and 0.75.
  • The χ² tests used the 165 examples on which the coders agreed for both functional features:
    • No example was both conscious and neg-TC.
    • DONT was used for 103 "unaware" cases and only 3 "aware" cases.
    • neg-TC split 32 aware and 27 unaware.
  • In DRAFTER, the coded examples trained a C4.5 tree that maps functional features to form. However, users still set these features by hand, because they are too hard to derive automatically.

Interesting ideas

  • I1. Harvest candidates with cheap string probes, cap each form at about 100 random samples, and then manually prune false hits such as conditional "if you don't" or affirmative "make sure"; this balances frequent and rare variants.
  • I2. Report chance-corrected agreement (Kappa) rather than raw percentage; a feature that looked acceptable at 76% agreement proved only moderately reliable.
  • I3. Run correlation analysis only on items where all coders agreed, so that noisy labels do not blur the form–function signal.
  • I4. Model the speaker's beliefs about the hearer (does he intend the act, does he know it is harmful) and use them to pick the surface form of a prohibition.
  • I5. The form of a prohibition can be read backwards as evidence about the writer's assumptions: a neg-TC warning implies the act would be accidental or that its harm is already known.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Build balanced evaluation sets of rare constructions this way, rather than letting very common patterns dominate the test data.
  • I2 (learn): Score agreement between witness parsers or LLM teacher tiers with Kappa per label, so that labels which agree only by chance are not trusted as silver data.
  • I3 (learn): Restrict training of repair rules or classifiers to the agreed-upon subset of silver data, and keep the disputed items as a separate diagnostic set.
  • I4 (back): When generating a warning from the meaning graph, choose "Don't X" for deliberate acts with unknown risks and "Be careful not to X" for accidental slips.
  • I5 (frames): Tag prohibitive speech acts with these inferred reader-model attributes, so that later implicature or QA steps can explain why a warning was phrased the way it was.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Carnegie Mellon UniversityPittsburgh · United States539 works · 179,533 citations🎓University of BrightonBrighton · United Kingdom14 works · 2,860 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Arts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksLiterature and Literary Theory · 519 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,263 works

Cite

Cite as Keith Vander Linden and Barbara Di Eugenio. 1996. A Corpus Study of Negative Imperatives in Natural Language Instructions. COLING 1996 Volume 1: The 16th International Conference on Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{linden1996corpus,
  title = {{A Corpus Study of Negative Imperatives in Natural Language Instructions}},
  author = {Keith Vander Linden and Barbara {Di Eugenio}},
  booktitle = {COLING 1996 Volume 1: The 16th International Conference on Computational Linguistics},
  year = {1996},
  doi = {10.3115/992628.992689},
  url = {https://doi.org/10.3115/992628.992689}
}

References

18 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Richard Alterman, Roland Zito-Wolf, and Tamitha Carpenter. 1991. Interaction, Comprehension, and Instruction Usage. Technical Report CS-91-161, Dept. of Computer Science, Center R)r Complex Systems, Brandeis Uniw~.r - sity. ↗
  2. Daniel Ansari. 1995. Deriving Procedural and Warning Instructions from Device and Environment Models. Master's tt,esis, University of Toronto. ↗
  3. Jean Carletta. 1996. Assessing agreement o~, classification tasks: the kappa statistic. Computational Lingustics, 22(2). find
  4. David Chapman. 1991. Vision, Instruction and Action. Cambridge: MIT Press. find
  5. Eirlys Davies. 1986. The English Imperative. Croom Helm. find
  6. Barbara Di Eugenio. 1993. Understanding Natural Language Instructions: a Computational Approach to Purpose Ulauses. Ph.D. thesis, University of Pennsylvania, December. Technical Report MS-CIS-93-91 (Also Institute for Research in Cognitive Science report IRCS-93-59,). find
  7. Laurence Horn. 1989. A Natural History of Negation. The University of Chicago Press. find
  8. J. McGowan and editors R. DuBern. 1991. Home Repair. London: Dorlin Kingersley Ltd. find
  9. Kathleen R. McKeown, Michael Elhadad, Yumiko l%lkumoto, Jong Lira, Christine Lombardi, Jacques Robin, and ~'ank Smadja. 1990. Natural language generation in COMET. In Robert Dale, Chris Mellish, and Michael Zoek, editors, Current Research in Natural Language Gene~ution, chapter 5. Academic Press. find
  10. Chris Mellish and Roger Evans. 1989. Natural language generation from plans. Computational Linguistics, 15(4):233 249, December. find
  11. Cdcile Paris and Keith Vander Linden. 1996. Drafter: An interactive support tool for writ-lug multilingual instructions. IEEE Computer. to appear. find
  12. C6cile Paris, Keith Vander Linden, Markus Fischer, A~,thony Hartley, Lyn Pemberton, R.ichard Power, and Donia Scott. 1995. A support tool for writing multilingual instructions. In Proceedings of the kburtcenth International Joint Conference on Artificial Intelligence, August 20 25, Montr6al, Canada, pages 1398 1404. Also availal)le as ITRI report [Tt{]-95-11. find
  13. J. Ross Quinlan. 1993. C/t.5: Programs for Machine Learning. Morgan Kaufmann. :1991. Reader's Digest New Comt)lete Do-It-Yourself Manual. find
  14. T. Rietveld and R. van Hout. 1993. Statistical Techniques for the Study of Language and Language Bchaviour. Mouton de Gruyter. find
  15. Sidney Siegel and N. John Castellan, Jr. 1988. Nonparametric statistics fl)r the behavioral sciences. McGraw Hill. find
  16. Keith Vander Linden and Barbara Di Eugenio. 1996. Learning micro-planning rules tbr preventative expressions. In Proceedings of the Eighth International Workshop on Natural Language Generation, Herstmonceux, England, 13 — 15 June 1996, June. To appear. find
  17. Keith Vander Linden and James Martin. 1995. Expressing Local Rhetorical Relations in hl-structional Text. Computational Linguistics, 21(1):29 57. find
  18. Stevcn Vere and Timothy Bickmore. 1990. A Basic Agent. Computational Intelligence, 6:41 60. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/linden-1996-corpus-study-negative-imperatives-natural/